ECTI Transactions on Computer and Information Technology (ECTI-CIT)

HAF-Net: Hybrid Attention Fusion Network for Diabetic Retinopathy Classification

2026-06-12T11:14:49+07:00

Diabetic retinopathy (DR) is a serious diabetes complication that can cause blindness if not treated promptly. Accurate DR classification plays an important role in supporting diagnosis. To improve classification performance, this research proposed a hybrid attention fusion network (HAF-Net) for DR classification. The HAF-Net integrates a convolutional operator with a convolutional block attention module (CBAM) and a bottleneck attention module (BAM) due to their advantages in enhancing feature representation, focusing on both local and global contextual information. These strengths allow the network to improve its ability to distinguish between subtle differences and help gather important features in medical images. A residual layer is also added to compensate for feature loss during training. A hybrid fusion of multi-layer perceptron (MLP), long short-term memory (LSTM), and bidirectional LSTM (BiLSTM) features are employed to achieve DR classification. The HAF-Net model is tested on a retinal fundus image dataset published on Kaggle for DR classification research, covering 11 disease types. The proposed HAF-Net achieved a training accuracy of 99.35% with a training loss of 0.0129. The hybrid model-MLP achieved the highest validation accuracy of 76.41% and test accuracy of 80.17%, demonstrating the best generalization performance. The results indicate that the HAF-Net achieves competitive classification performance compared with previous studies and represents a promising approach for automated DR classification.

Horned Lizard Optimization-Driven CA-InceptionResNet Framework for Enhanced Dementia Diagnosis

2026-06-23T09:59:05+07:00

Dementia is a progressive neurological disorder affecting millions worldwide. It places a tremendous burden on global healthcare systems and research. Early detection and accurate diagnosis are essential for effective disease management. The purpose of this paper is to present an integrated framework using Deep Learning (DL) to help improve dementia diagnosis. The framework uses an Adaptive Difference of Gaussian (ADoG) filter to sharpen important edges on neuroimaging scans while reducing background noise, producing sharper (more focused) data in neuroimaging scans that contain significant amounts of data. To segment brain tissue into anatomical regions affected by neurodegenerative disorders using the Hybrid Agglomerative Clustering Refinement (HACR) technique. The segmented anatomical regions are analyzed using a polar-coordinate-based Gray-Level Co-Occurrence Matrix (GLCM) to extract rotation-invariant texture features, ensuring robust and orientation-independent characterization of tissue patterns. For a classification stage, proposed framework utilizes a Channel-Attention Inception-ResNet (CA-InceptionResNet) model. By integrating feature extraction stength of InceptionResNet architecture with channel-attention mechanism, proposed model effectively captures discriminative features and improves classification of dimentia sublevels. For classifier performance optimization, the Horned Lizard Optimization (HLO) algorithm is employed to optimize the model's hyperparameters. The proposed model is implemented in Python and achieved an accuracy of 98.98%,precision of 99.12%, recall of 98.96%, F-score of 99.04% and specificity of 99.58%, demonstrating its effectiveness in accurate and reliable classification.

Beyond Mean Aggregation: A Variance-Enhanced Graph Neural Network for Android Malware Detection

2026-07-30T15:03:35+07:00

Android malware continues to pose a significant threat to mobile security, with millions of malicious applications detected annually. Graph Neural Networks (GNNs) applied to Function Call Graphs (FCGs) have emerged as a promising approach for malware detection. However, existing GNN-based methods rely predominantly on mean aggregation, capturing only the first-order statistical moment of neighborhood representations—insufficient for heterogeneous FCGs where malware classes exhibit fundamentally different connectivity patterns. We present an empirical analysis demonstrating that neighborhood degree variance differs significantly across malware classes (ANOVA, p < 0.001), motivating variance as an explicit aggregation signal. Based on this finding, we propose VE-GNN, a Variance-Enhanced Graph Neural Network incorporating mean and variance aggregation within a residual message-passing framework. Experiments on MalNet-Tiny with five independent runs show that VE-GNN achieves a macro F1-score of 0.9261 ± 0.0077, outperforming GIN by 4.57% (paired t-test, p < 0.001). Ablation studies confirm that mean-variance aggregation is optimal, as higher-order moments degrade performance without adaptive selection. The proposed method offers an effective and computationally efficient solution for GNN-based Android malware detection.

Formal Verification of Ethical Behavior in Mental Health Conversational Systems Using Finite-State Models

2026-07-30T14:52:42+07:00

The evaluation of mental health conversational systems has relied on benchmark-driven and performance-based metrics, which are limited in their ability to guarantee ethical safety. Although conversational systems achieve high performance scores, critical ethical failures may occur in contexts involving emotional vulnerability, medical requests, or potential selfharm. To address this limitation, this study proposes a formal verification framework based on finite-state machine (FSM) model checking to assess ethical behavior. The proposed framework models multi-turn conversations as nite execution traces and uses expert-driven risk taxonomies to define ethical safety requirements. Ethical requirements are applied as accepted transitions and forbidden transitions, enabling deterministic verification of conversational trajectories. The approach offers reproducible and platform-independent ethical assurance based on formal methods. The study evaluated multiple conversational AI platforms using fixed-length sessions and a balanced set of risk scenarios. The difference results across platforms show that ethically unsafe trajectories can still emerge despite controlled experimental conditions. The ethical compliance rate, risk-level violation rate, and forbidden transition frequency successfully expose behavioral failure modes that remain hidden under conventional evaluation paradigms. The study supports a shift from performance-oriented evaluation toward verification-driven ethical assurance. In addition, FSM-based verification is demonstrated in promise for safety-critical conversational AI systems.

Real-Time Frailty Detection for Elderly Care in Indoor Environments

2026-06-12T11:34:52+07:00

The rapid growth of the global older population has become a major public health concern, placing increasing pressure on healthcare systems to address age-related health condition effectively. Frailty syndrome is an age-related condition characterized by increased vulnerability due to reduced physiological reserves, leading to impaired gait, balance, and muscle weakness. It also causes functional decline, which elevates the risk of falls, disability, hospitalization, and rapid progression to dependency. Early identification of frailty is crucial for enabling timely preventive interventions and improving health outcomes and quality of life among elderly individuals. To address this challenge, this study presents a machine learning-based frailty detection system that enables continuous and remote monitoring of gait characteristics in indoor environments. The proposed system is integrated with an Ultra-Wideband (UWB) indoor positioning system to provide precise localization. It employs a wearable UWB tag equipped with a tri-axial accelerometer for data collection. Data stream processing techniques are applied to perform real-time gait analysis and frailty detection from acceleration data patterns. Experimental results demonstrate that the proposed system is practical for real-world deployment. Among the evaluated models, the Random Forest model achieved the best overall predictive performance, with an overall classification accuracy of approximately 87.32% and an average false positive rate of 6.3%.

Explainable Deep Fusion for Breast Cancer Classification in Histopathology Images

2026-07-09T10:50:56+07:00

Breast cancer diagnosis from histopathology images is accurate but labor-intensive and subject to interobserver variability. Deep learning has shown strong performance on digital pathology benchmarks; however, limited transparency, magnification variability, and patient-level data leakage can reduce clinical trust and robustness. This paper proposes an explainable deep fusion framework for magnification-robust breast histopathology classification on the BreakHis dataset. In the experiments, all folds are generated at the patient level, and model selection is performed in two stages: first, single backbones and backbone-pair fusion models are compared without attention; second, attention modules are evaluated on the selected fusion pair. The results identify VGG16ResNet18 fusion without attention as the nal configuration, achieving an accuracy of 0.8890, an F1-score of 0.8888, and an AUC of 0.9326 under mixed-magnication, patient-level 5-fold cross-validation. Although the improvement over the strongest single-backbone baseline is modest and not statistically significant in the fold-level Wilcoxon test, an image-level discordant-case analysis shows that the proposed fusion model corrects more EfficientNetV2 errors than it introduces. To support transparency, Grad-CAM-based case analysis is used to examine corrected, newly misclassified, and persistently misclassified examples, with LIME and SHAP retained as complementary explanation tools when generated for selected cases. These results suggest that feature-level fusion can provide useful gains for some difficult histopathology images while highlighting remaining limitations caused by ambiguous tissue morphology, staining variation, magnification-dependent texture changes, and attention to non-diagnostic regions.

Latency-Aware Cross-Domain Rice Variety Classification

2026-08-13T09:45:07+07:00

Accurate rice variety classification is essential for quality control, seed authentication, and agricultural supply chain management. Although deep learning methods have achieved high classification accuracy, most existing studies overlook domain shifts, deployment efficiency, and model interpretability. This paper proposes a latency-aware cross-domain rice variety classification framework that integrates proxy-domain grouping, GroupK-Fold evaluation, pretrained CNN feature extractors, PCA-based dimensionality reduction, classical machine learning classiers, and explainability analysis. A sensitivity analysis was conducted to determine the optimal proxy-domain configuration and feature dimensionality. Experimental results across four publicly available rice datasets showed that K = 5 provided the most suitable proxy-domain structure for domain-aware validation, while PCA with 150 components retained approximately 99.08% of the original variance and achieved the best cross-domain performance. Among the evaluated backbones, ResNet50 with partial fine-tuning consistently produced the strongest robustness under domain shifts, whereas EfficientNetB0 frequently achieved the highest official test performance. The proposed framework attained weighted F1-scores of up to 0.999 on the official test sets, with 95% bootstrap confidence intervals confirming result reliability. Latency benchmarking across CPU, GPU, and TensorFlow Lite environments revealed that TensorFlow Lite reduced inference latency from approximately 150280 ms to 58 ms per image, supporting efficient edge deployment. Furthermore, Pareto frontier analysis identified MobileNetV3Small as a highly competitive deployment-oriented backbone, while Grad-CAM analysis indicated that most remaining classification errors were associated with visually similar rice varieties. Overall, the proposed framework provides a robust, efficient, and interpretable solution for rice variety classification under realistic domain-shift conditions.

Vision Transformer Architecture for Simultaneous Detection and Severity Grading of Diabetic Retinopathy and Diabetic Macular Edema

2026-08-20T10:20:08+07:00

Diabetic retinopathy (DR) and diabetic macular edema (DME) are two major causes of preventable blindness. Despite their great success in DR and DME detection or grading individually, few methods have explored automated simultaneous grading of DR and DME from fundus photographs in clinic settings. In this work, we introduce DualHead-ViT, a multi-task framework with a pretrained ViT-B/16 backbone and a novel cross-task attention fusion module for jointly grading the severity of DR and DME. We project two separate CLS token vectors for DR and DME, respectively, and design a cross-task attention module to interact features of DR and DME bidirectionally. We validated our model using the Indian Diabetic Retinopathy Image Dataset (IDRID), employing stratified 5-fold cross- validation alongside CLAHE-preprocessed images and a class-weighted loss with ordinal regularization. Across 5 levels of DR severity, we report mean quadratic kappa of 0.802 ±0.010 and AUC of 0.831 ±0.018, and across 3 categories of DME risk, we report kappa of 0.759 ±0.067 and AUC of 0.895 ±0.026. Our findings support the feasibility of joint transformer-based grading for DR and DME using a single computational architecture.

Secure Communications in RIS-Aided NOMA Networks Joint Active and Passive Beamforming with Physical Layer Security

2026-07-06T15:57:42+07:00

In this paper, we study the physical layer security of reconfigurable intelligent surface (RIS) aided non-orthogonal multiple access (NOMA) downlink systems. A multi-antenna base station (BS) serves two NOMA users in the presence of a passive eavesdropper, and a RIS equipped with N passive phase-shifting elements is employed to assist the communication. To enhance the sum secrecy rate, we propose a joint active and passive beamforming optimization framework for secrecy maximization based on alternating optimization (AO). The active beamformer is designed with the maximum ratio transmission (MRT) beamformer towards the user with stronger channel, and the passive RIS phase shifts are optimised by an element-wise block coordinate descent method with one-dimensional grid search. Simulation results show that the proposed scheme significantly outperforms the following three benchmark schemes, i.e., random phase RIS-NOMA, orthogonal multiple access with RIS, and conventional NOMA without RIS, over a large range of operating conditions. In particular, the proposed algorithm is shown to achieve the sum secrecy rate of 16.6 bps/Hz at 40 dB SNR, convergence within six iterations, and secrecy outage probability close to zero for all the considered SNR values. In addition, the improvements brought by more RIS elements, BS antennas, and eavesdropper distance are also confirmed by extensive Monte Carlo simulations.

Adaptive Hybrid Stacking Ensemble Learning for Real-Time Forest Fire Risk Prediction

2026-08-20T10:27:32+07:00

Forest and land fires remain a recurring environmental issue, particularly in regions with high climate variability such as Indonesia. This study proposes an Adaptive Hybrid Stacking Ensemble approach for real-time fire risk prediction by integrating meteorological data and IoT-based sensing systems. The model combines multiple base learners with a rotation strategy and an exponential accuracy weighting mechanism to optimize the contribution of each model. A dataset consisting of 4,931 meteorological records was used, including variables such as Average Temperature, Relative Humidity, Rainfall, Wind Speed, and Soil Surface Moisture. Experimental results show that the proposed model significantly outperforms conventional stacking, achieving an accuracy of up to 92.47%, compared to 86%87% in baseline stacking. The optimal performance was obtained at α= 5 in the weighting scheme, which effectively balances model contribution. Evaluation using confusion matrix, ROC curve, and K-fold cross-validation confirms the model's robustness, with AUC values reaching 0.981.00 and stable accuracy across folds. In addition, Explainable AI using SHAP reveals that Rainfall and Soil Surface Moisture are the most influential features in fire risk prediction. The integration with IoT enables real-time data processing and prediction, making the system applicable for early warning and disaster mitigation.